Context Engineering

Plan injection can make an actor follow adversarial instructions while producing reasoning that looks benign to a monitor. Agent builders should treat chain-of-thought inspection as one signal, not a security boundary.

Gavel routes skills from a frozen model’s internal states without loading skill text into context. It suggests large skill libraries may need model-native selection rather than metadata-heavy prompts.

AI did not start web-design sameness, but it accelerates it. Builders should give design agents explicit brand context, varied references, and checks for repetition, fit, and coherence.

ConvMem replaces sequential long-context memory updates with a parallel hierarchy of summaries, offering a training-free pattern for agents that must reason over oversized inputs.

An MCP widget can reduce an agent’s useful work if rendering and retrieval share one tool. Keep data exploration separate, then render only the results the model selects.

In one same-prompt test, a context engine nearly halved agent token use. The design goal is targeted, permission-aware retrieval that resolves conflicts across engineering systems.

Shared agent memory needs user-scoped reads, human-approved writes, and per-request credentials. The talk offers a concrete security boundary for company-wide context and tools.

Cross-silo agents are primarily a context and privacy-boundary problem. Start with low-sensitivity data, explicit sharing policy, and approval at the moment information leaves a silo.

A telecom diagnosis pipeline structures telemetry before forcing a decision path and evidence-grounded explanation. The pattern is relevant to agents reasoning over heterogeneous operational data.

For fuzzy requests, an agent should identify the missing constraint with the most decision value, elicit it in the right modality, then choose a response format suited to the task.

Shared agent skills can turn team know-how into portable workflows, but unmanaged catalogs create duplication and policy risk. Treat skills as versioned, evaluated dependencies.

Agents often stop after finding plausible code context. Werry argues for task-specific retrieval that also surfaces decisions, conventions, PR history, and expertise before planning or review.

CritICL turns recurring failures from smaller models into critique examples for stronger models, aiming to improve inference with fewer generations and lower token cost than repeated sampling.

For fast-changing libraries, ship concise agent-facing docs inside the package as well as on the web; agents may inspect installed code before visiting your site.

Notion treats GTM automation as a shared context system for humans and agents, not an AI layer over disconnected tools. The key design choice is keeping risky customer actions human-approved.

Tool-by-tool restrictions failed to contain a chip-design agent. Enforce permissions at the substrate, isolate files by role, and propagate approved spec changes from one source of truth.

LLM compliance generation behaves differently under vague and strict schemas: vague artifacts need richer context, while rigid formats can stay consistent yet hallucinate.

A coding agent with repository access still recommended a known outage trigger. Background agents need reconciled decisions and incident context, not just more raw MCP results.

ConceptTS turns LLM-proposed concepts into executable labels and interpretable forecasting bottlenecks, enabling concept-level inspection and intervention.

IAR turns a fixed document corpus into parametric knowledge through separate injection, QA alignment, and recovery stages. It improved domain QA while retaining more general capability than vanilla SFT.

Lease End replaced a fine-tuned intent classifier with skills and runtime context, cutting production fixes from about a week to under an hour. Higher API spend was offset by lower maintenance cost.

Automated code review depends less on a stronger model than on codifying team rules, architecture, incidents, and service contracts into context both humans and agents can inspect.

Nereu argues that coding agents need an engine-native vocabulary, not more generated glue code: tags, reusable systems, and selective scene context turn game intent into reliable actions.

In this AI-tutor evaluation, keeping full history beat compaction on recall and cost because cached tokens were cheap. Measure your constraint before shortening agent context.

Agent handoffs should preserve task-relevant predictive state, not blindly compress the transcript. The paper offers a three-part record for carrying decisions, evidence, and irreducible observations across sessions.

SheetCompass keeps spreadsheet structure as a hierarchical relation graph instead of flattening cells into text. Its agent memory retains task-relevant context across sheets, tables, and columns.

A small company-enrichment test suggests rented web context is convenient for changing queries, while repeated stable queries may justify owning the pipeline. The reported crossover was around 15,000 queries.

Persistent memory is a compute and product tradeoff, not just retrieval. Profiles need conflict detection, visibility, editing, and deliberate update cadence before agents can rely on them.

A useful agent knowledge base starts with abundant raw Markdown, then adds controlled tags, backlinks, and generated indexes. Scheduled agents can maintain it, but sync and taxonomy need explicit rules.

Stronger models can still behave like smart novices inside company-specific workflows. Agents need learning loops that compress experience into reusable context, procedures, and judgment without losing reliability.

Training directly on a private corpus can memorize it without producing useful behavior. Durable context learning likely needs synthetic study tasks and recursive self-improvement, not a single fine-tuning pass.

Anthropic’s harness fixes for Sonnet 4.5 became harmful under Opus 4.5. Treat model-specific scaffolding as replaceable, and keep sessions durable enough to recover context and tool failures.

Codex exposes several reusable harness patterns: deferred tools, persistent browser execution, sandboxed edits, automated permission review, WebSockets, and server-side compaction.

Faster code generation can overwhelm merge queues and scatter team attention. The talk proposes durable shared plans as agent state, shifting human effort toward decisions and prioritization.

A walkthrough of Anthropic’s architect exam doubles as a harness checklist: inspect stop reasons, narrow subagent scope, isolate outputs, and compact long-running context.

Blast Radius predicts which code and context a prompt will reach, then reversibly archives unused history. Across seven OpenAI models, it cut token consumption by 17–26%.

Controlled LLM rewriting made harder financial sentences cheaper to process with DisCoCat, cutting circuit size by over 70%, but downstream accuracy improved only modestly.

MIST tests whether models use good context while resisting bad context, exposing agents that appear robust only because they ignore external evidence altogether.

A Greek RAG study shows generic multilingual embeddings can lose to BM25 in specialist domains, while targeted adaptation sharply improves retrieval and grounded answers. Language-specific evaluation still matters.

Chained RLM repeatedly gives the same model a fresh context while preserving summaries, a blackboard, and artifacts—a concrete harness pattern for limiting error buildup.